Flash-BoN: A Stronger Baseline for Diffusion Inference and Post-Training

RisingSayak introduced a new paper and code focused on Flash-BoN for diffusion model inference and its extension to post-training, Flash-Flow-GRPO. The core argument is not just about offering a faster method, but recalibrating the evaluation metrics for test-time scaling: when verifier costs and actual wall-clock time are factored in, a well-executed best-of-N (BoN) remains a highly competitive baseline, with Flash-BoN pushing this efficiency even further.

Key Details and Evaluation Metrics

RisingSayak argues that looking only at NFE creates an illusion of efficiency, as many methods exclude the overhead of additional verification or reranking. The selling point of Flash-BoN is low-cost candidate generation: expanding the candidate pool while striving not to sacrifice diversity. He emphasizes that diversity is crucial for test-time scaling to genuinely expand the search space; highly similar candidates yield diminishing returns despite increased numbers.

Results and Applicability

According to RisingSayak, Flash-BoN's advantages are not limited to a single benchmark or hyperparameter tuning, but hold true across multiple benchmarks. When wall-clock time is included in the comparison, other baselines lag significantly, a conclusion consistent across different model sizes and families. For image generation tasks, they explored how to utilize inference-time scaling under a fixed time budget (e.g., 60 seconds), noting that repeated intermediate verification is more costly than intuitively expected.

Extension to Post-Training

The same "cheap candidate generation" approach has been applied to RL post-training. RisingSayak shared that Flash-Flow-GRPO can match the performance of the baseline Flow-GRPO with 10x fewer steps. This frames Flash-BoN not merely as an inference optimization trick, but as a unified pipeline for both inference and post-training efficiency.

2026-07-14 ~ 2026-07-14 · 7 related posts